henhao
Sign in
GenomicsbioRxiv
Preprint

A pangenome-graph approach for mapping and imputing barley sequences

Joan SàrriaiD, Hamza Amhal, C. J. Ramírez, Ernesto IgartuaiD, Ana Maria CasasiD, Bruno Contreras‐MoreiraiD

Abstract

Barley (Hordeum vulgare) is a key cereal crop with exceptional adaptation to diverse environments. With a large, highly repetitive diploid genome, barley presents challenges for pangenome representation. Starting from the reference genome MorexV3, we describe the construction of a barley graph (Pan20) representing the global diversity of landraces and cultivars captured in the public pangenome V1. For mapping arbitrary sequences, a greedy strategy is proposed that combines GMAP alignment followed by intersection with a Practical Haplotype Graph (PHG). This enables presence-absence variation detection and provides a consistent MorexV3 physical coordinate system across genotypes, enabling comparative analysis and visualization. For imputation of genomic data, the PHG approach relies on k-mer pseudo-alignment against the graph. Benchmarks show that Pan20 can accurately align barley genomic and transcriptomic sequences, including those not present in the Morex reference, revealing that a third of long genomic sequences map on non-reference genomes. Moreover, experiments with Genotyping by Sequencing and low-pass sequencing data indicate that FASTQ files can be efficiently mapped and imputed against the graph, preserving local haplotype context. This flexible and scalable graph framework allows barley researchers to explore genetic diversity beyond a single reference and facilitates analysis of diversity panels at the haplotype level, going beyond SNPs. Documentation and a Docker container are available at https://github.com/eead-csic-compbio/barleygraph. The graph sequence mapping utility was added to the Web application https://barleymap.eead.csic.es.

PDFDOICodeData
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
0
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (10 Aug 2026) · CC BY 4.0

Pan20, a Practical Haplotype Graph of 20 barley genomes, assigns sequences MorexV3 coordinates and imputes genotyping-by-sequencing (GBS) reads at 0.06% base-call error versus 0.32% for variant calling.

Barley pangenome V1 provides 20 annotated genomes, but each has its own coordinates, gene identifiers and annotation, and graph-building tools risk oversized graphs or collapsed repeats in this large repetitive genome. Building on PHGv2 (already used in wheat and maize), this work adds a greedy GMAP scan across genomes so that sequences overlapping graph ranges, including ones missing from MorexV3, receive a shared MorexV3-based coordinate, and packages PHG k-mer imputation of FASTQ reads against the graph.

Key results

  1. ShownPan20 maps sequences beyond the reference. A few percent of transcripts, and a third of HiFi genomic reads, failed the MorexV3 cut-offs but matched another Pan20 genome. Nearly all mapped sequences still received MorexV3-based graph coordinates.

    Identity ≥98%, coverage ≥95%. PanBaRT20 (79,580): 99.9% mapped, 97.9%/98.4% (mmap_pro/gmap_geno) given graph coordinates, up to 2% non-reference. pantranscriptome16 (47,588): 95.5%, 91.2%/90.1%, up to 7%. 10K short reads (LGDiablo): just over 93% mapped, 10% non-reference. 10K HiFi reads (HOR_495): 81%, 34%. · Mapping genomic fragments and transcripts; Transcriptomic and genomic sequences; Figure legends

  2. ShownGraph imputation of reads from accessions outside Pan20 raised mapping rates. Scored against each accession's own assembly, it also lowered overall base-call error (mostly SNPs) compared with standard variant calling on MorexV3. Indel errors were higher with imputation.

    25 GBS accessions, mmap_pro graph: reads mapped 97.71% to Pan20 vs 94.71% to MorexV3 vs 97.78% to own assembly; error 0.06% vs 0.32% (standard calls, depth ≥10); indels alone 3.82% vs 3.09%. Low-pass whole-genome sequencing, 7 genotypes: 97.73% vs 97.45% mapped; error 0.06% vs 0.36%; indels 2.61% vs 1.67%. · Graph-based imputation of GBS and low-pass data; Imputation tests; Figure legends

  3. ShownOn pangene coding sequences (CDS) with known positions, the PHG-default mmap_pro graph gave the most exact coordinates, and combined GMAP plus minimap2 range refinement at the default identity was the best setting. gmap_geno aligned slightly more sequences but less precisely.

    mmap_pro: 97.1% exact starts and 96.1% exact ends. gmap_geno at 100% identity returned more alignments but only 83.9% exact starts. Offsets of a few hundred bp covered 99% of mmap_pro results across genomes. · Mapping pangenes with known coordinates; Barley pangenes

  4. ShownIn flowering and spike-architecture genes, no single graph-building mode handled every structural variant. Proali graphs failed at HvCEN in genomes carrying a chr2H inversion; genoAli graphs failed at Vrs1 in OUN333. Copy-number-variable CBF transcripts collapsed onto fewer loci.

    22 of 28 loci mapped in all genomes by at least one of four graphs; 6 likely accessory. The HvCEN inversion is shared by 7 Pan20 genomes. VRN-H2, absent from MorexV3, was resolved, showing presence-absence among non-reference genomes and up to three copies in Akashinriki. Three Igri CBF transcripts landed on two Pan20 loci. · A close look at flowering pangenes; Mapping genomic fragments and transcripts; Tables

  5. ShownA 20-genome barley PHG can be built on one conventional server and stays compact. Alignment and imputation indices add substantial disk and memory.

    mmap_pro graph: 16,441 core ranges, 47,672 accessory (2 to 19 genomes) and 173 unique. Built in a week on a Linux server with 256GB RAM. Graph about 4GB; GMAP indices 150GB; imputation indices need 120GB free disk (13GB final files); phg rope-bwt-index takes up to ~250gb RAM. · Building Pan20 pangenome graphs; Barleygraph; Discussion and conclusions

Methods, data and reproducibility

Four Pan20 graphs were built with PHGv2 from MorexV3 plus the other pangenome V1 assemblies, varying the liftover aligner (minimap2 or GMAP) and the AnchorWave mode (proali or genoAli). Barleygraph scans input sequences with GMAP genome by genome until one passes the identity and coverage cut-offs, intersects the hit with graph ranges, and imputes FASTQ reads via PHG k-mer mapping and Viterbi path finding.

  • Barley pangenome V1 genome assemblies and gene annotations (MorexV3 reference plus landraces, cultivars and one wild genotype) used to build Pan20 · Hordeum vulgare (barley) · n = 20 genomes
  • PanBaRT20 transcripts from the 20 Pan20 genotypes (first annotated isoform per gene) · Hordeum vulgare (barley) · n = 79,580 gene models
  • pantranscriptome16 long-read and de novo assembled transcripts (longest isoform per cluster) · Hordeum vulgare (cultivated and wild barley) · n = 47,588 transcript clusters
  • Whole-genome short reads of cultivar LGDiablo (DNBSEQ-T7, 2×150bp); first ten thousand sequences used in the mapping benchmark · Hordeum vulgare (barley) · n = 10K sequences
  • HiFi long reads of landrace HOR_495; first ten thousand sequences used in the mapping benchmark · Hordeum vulgare (barley) · n = 10K sequences · ERR10661231 (ENA)
  • Precomputed barley pangene CDS (release 19052025; pangenes in ≥17 Pan20 genomes, HOR_10350 excluded), including a curated flowering and spike-architecture subset · Hordeum vulgare (barley) · n = 28 flowering-related pangenes in the curated subset; total not stated
  • GBS FASTQ files and genome assemblies of barley accessions not in Pan20 (ENA and IPK data archive) · Hordeum vulgare (barley) · n = 25 accessions
  • Low-pass whole-genome sequencing reads of barley cultivars · Hordeum vulgare (barley) · n = 7 cultivars
Methods & tools
PHGv2, AnchorWave, minimap2, GMAP, AGC, bedtools, IQ-TREE, bwa, samtools, bcftools, phgtools, BARLEYMAP
Code
https://github.com/eead-csic-compbio/barleygraphhttps://github.com/eead-csic-compbio/barleygraph/pkgs/container/barleygraphhttps://github.com/jsarriaa/PHGv2Toolshttps://github.com/Cantalapiedra/barleymap_webhttps://barleymap.eead.csic.es/
Data & accessions
https://github.com/eead-csic-compbio/barleygraph/releases/tag/20260807https://floresta.eead.csic.es/plant-pan-genomes/mappinghttps://ics.hutton.ac.uk/panbart20https://github.com/eead-csic-compbio/barley_pangenes/releases/tag/19052025https://github.com/eead-csic-compbio/barleygraph/blob/main/paper_Pan20/flowering_bench_Pan20.tsvhttps://galaxy-web.ipk-gatersleben.de/libraries/folders/Fd071e794759ab192
Software & environment
Docker container with a setup_graph script that downloads precomputed graphs (e.g. -G Pan20-mmap-pro; -g also builds GMAP indexes). PHGv2 release 2.4.74.229, minimap2 2.24-r1122, GMAP 2013-09-30, AnchorWave v1.2.5, IQTREE v1.6.1.
Assay conditions & compute
Graph build took a week on a Linux server with 256GB RAM. Graph about 4GB; GMAP indices 150GB; imputation indices need 120GB free disk, 13GB final files and >12GB RAM. Indexing with phg rope-bwt-index takes up to ~250gb RAM.
Reproducibility gaps
Input genome, GBS and low-pass sample lists are only in supplementary tables (S1, S5, S6) not included in the text. No accession is given for the full LGDiablo read set, only a shared 10K-sequence subset. The haplotype analysis mode is distributed only as a Docker container.
Limitations (6)
  • The reference keeps a founder effect: every genome is aligned to MorexV3, and its gene annotation defines the primary set of ranges. · stated by the authors
  • No single liftover or alignment mode resolved all structural variants (HvCEN inversion vs Vrs1 in OUN333), so several graph-building strategies need testing. · stated by the authors
  • The greedy scan can misplace copy-number-variable loci (the CBF example): the result may depend on genome scanning order, and scanning MorexV3 alone affects any locus where MorexV3 has fewer copies. · stated by the authors
  • Imputation is expected to degrade for samples whose reads cover the genome unevenly, and for haplotypes absent from the graph, which the Viterbi path replaces with the most similar represented haplotype. · stated by the authors
  • Imputation was compared only with direct variant calling on MorexV3. No linkage-disequilibrium (LD) imputation tool such as Beagle was benchmarked, and indel error was higher with PHG imputation (3.82% vs 3.09% for GBS). · noted in this summary
  • Genomic-read mapping rests on 10K sequences from one short-read and one HiFi sample. The pangene CDS benchmark uses sequences from the same genomes that built the graph. · noted in this summary

Who should read:Read this if you work on barley or another large crop pangenome and need consistent reference coordinates for sequences across genomes, or want to impute GBS or low-pass genotypes from a haplotype graph.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISBradbury '22Muñoz-Amatr '15Jordan '21Franco '20Sallam '24Long '21
65% semantic matchBioinformatics
Peter J. Bradbury, Terry M. Casstevens, Sarah E. Jensen+6 more·Bioinformatics·Jun 24, 2022· 61 citations